feat(planner): AutocompleteCache — 4096 interdependent attention heads 4 modules, 20 tests: cache/kv_bundle.rs: HeadPrint, AttentionMatrix (HIP 64×64, TWIG 256×256) cache/candidate_pool.rs: CandidatePool, Phase (Exposition→Coda), ranked emit cache/triple_model.rs: TripleModel (self/user/impact), Plasticity, DK, Truth cache/lane_eval.rs: LaneEvaluator, Tension (Euler-gamma noise floor) Next: refactor HeadPrint → palette u8 index with O(1) lookup, add SPO decomposition (3×u8), Pearl 2³ mask (8 projections), NARS inference rules as style weight vectors over 8 projections. https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK#73